Application of machine learning methods and Sentinel-2 data for multitemporal land-cover classification in conflict-affected areas

Abstract. In many regions of the world, especially those affected by armed conflicts, urbanization, or intensive environmental transformations, a high dynamic of land use and land cover changes is observed. Reliable monitoring of these processes requires the application of classification methods that ensure both high thematic accuracy and temporal consistency. This paper presents a multitemporal classification methodology based on Sentinel-2 optical data and machine learning models. 
The research was conducted for the city of Sievierodonetsk (Luhansk Oblast, Ukraine) – an area that suffered significant destruction in 2022 as a result of military operations. The aim of the analysis was to identify land cover changes in the years 2021-2025 using three classifiers: k-Nearest Neighbors (kNN), Random Forest (RF), and Gradient Boosting Classifier (GBC), combined into an ensemble system based on dynamic confidence weighting. 
Quality assessment using the recall metric showed that the fusion method outperformed individual classifiers, achieving average values of 0.87-0.96, while classical models obtained 0.81-0.89. The largest changes (39%) occurred in the years 2022-2023, coinciding with the period of greatest military activity. The proposed method achieved the highest classification quality indices (F1 = 0.93, Acc = 0.98 for 2021), surpassing global products and models based on AlphaEarth. In subsequent years, high stability was maintained (F1 ≥ 0.88), confirming the effectiveness and robustness of the approach under various environmental conditions.